Google Cloud has open-sourced a lightweight AI agent that rethinks how AI systems handle long-term memory, replacing the standard retrieval-augmented generation (RAG) pipeline with a continuously running process that reads, consolidates, and queries structured memories stored entirely in SQLite.
The Always-On Memory Agent, built with Google’s Agent Development Kit and running on Gemini 3.1 Flash-Lite, operates 24/7 as a background process. It uses no vector database and no embeddings. Instead, an LLM reads incoming content, extracts structured information, and periodically consolidates related memories, a design that its creators compare to how the human brain processes information during sleep.
“Most AI agents forget. They process a request, answer it, then drop the context,” the project’s documentation states.
The system uses three specialized subagents working under a single orchestrator. An IngestAgent extracts summaries, entities, topics, and importance scores from incoming content, supporting 27 file types including text, images, audio, video, and PDFs. A ConsolidateAgent runs on a timer, defaulting to every 30 minutes, that reviews unconsolidated memories, finds connections between them, and writes synthesized summaries and key insights. A QueryAgent answers questions by reading all stored memories and consolidation insights, citing the specific memory IDs it draws from.
The approach represents a fundamental departure from the dominant RAG pattern. In a traditional RAG system, documents are embedded into a vector store at ingest time and retrieved passively at query time. There is no active processing between ingest and retrieval. The Always-On Memory Agent, by contrast, processes memory actively during the gap between ingest and query, linking related information and generating cross-references that a vector similarity search would miss.
The agent is available on GitHub under Google Cloud’s generative AI repository. It watches a configurable inbox directory, consolidates automatically, and serves an HTTP API on port 8888 for ingest and query operations. A Streamlit dashboard provides a visual interface for browsing and managing memories.
Use cases range from research assistants that ingest PDFs, meeting audio, and screenshots over the course of a week, and can later connect a cost target to a reliability problem autonomously, to support agents that store past tickets as structured memories and answer new questions with cited references.
For developers and enterprises building AI systems that need to maintain context across sessions without the infrastructure overhead of a vector database, the agent offers a simpler alternative: one SQLite database, one background process, and a prompt that tells the LLM to think before it stores.
Sources: “Google Cloud’s Always-On Memory Agent Replaces RAG and Embeddings With Continuous LLM Consolidation on Gemini 3.1 Flash-Lite” (MarkTechPost, July 18, 2026); Google Cloud GitHub repository